Diversity must sit on dimensions the user cares about; cross-category mix does not fix a single viewpoint

What it is

Home has food, sport, and finance — category entropy is high — and the three news items still share one stance and one outlet. For someone here for public affairs, that is not diversity. Diversity on relevant dimensions means the spread has to land on the axis this visit cares about — stance, author, style, difficulty — not on a cheap axis (category, duration, language) used to tick a box.

A cross-category platter is the cheapest diversity metric. It often misses the dimension that is actually narrowing.

Why it happens

Dashboards like axes that can be tagged automatically: leaf category, duration bucket, language. Stance, narrative frame, creator networks cost more, so the optimiser fills coverage on the cheap axis and leaves the expensive one as a point. The user’s task defines “relevant”: looking for a recipe, cuisine is relevant and politics is not; reading news, the reverse. High coverage on an irrelevant axis comforting a point on the relevant one is metric substitution.

The core harm in Pariser’s bubble is issues and viewpoints locked in; it is not “whether a sports clip was spliced in.” The sports clip can even make the coverage report look better while hiding the real shrink. Pick the wrong dimension and structured methods will diligently spread in the wrong direction.

Studying it

Elicit the task: what is this visit comparing. Then add a quota on that dimension and on an irrelevant one. Measure: perceived diversity relevant to the task, task performance (can they see an opposing stance, a different author), coverage on the irrelevant axis, clicks. Independent variables: which dimension gets the quota, how users describe the axis they care about. Dependent variables: real spread on the relevant dimension, false judgment that an irrelevant platter is “already rich.”

Do not take category entropy as success in a news setting. That is the definition of fake diversity.

Where it stops holding

On leisure browsing with no single task, the relevant dimension is unstable; offer switchable axes (“another author,” “another style”) rather than pinning one. For children’s content the relevant dimension may be difficulty and topical safety, not stance. When the catalogue has only one value on the relevant dimension (a single-author podcast), diversity cannot be made; use an entry other than related recommendation. This entry chooses the dimension. It does not discuss random shuffle, and it does not discuss writing diversity as an objective.

Applying it

• For each main surface, write one sentence: “what is the user comparing on this visit.” Put quotas only on the axis in that sentence. If home hosts several tasks, split rails; do not serve them all with one category entropy.
• Split reports by relevant dimension: on a news surface, report outlet and stance, not only top-level category.
• Check: on users whose viewpoint has already narrowed, add only a cross-category quota. If category entropy rises and stance entropy does not, that is fake diversity. Move the quota onto outlet or stance and watch stance entropy.

Related

• Same group: L6.08.1 Narrowing is produced by the objective itself; no one has to intend it · L6.08.2 Diversity costs short-term clicks, so it has to be an independent objective · L6.08.3 Users cannot evaluate content they never saw; absence produces no visible feedback · L6.08.4 Random injection is not structured diversity; it only adds noise
• Nearby: L6.02 Filter Bubbles · L5.05 The Moderation Principle of Transparency · L6.12 Conflict Between Personalization and Predictability
• Search terms: diversity on relevant dimensions · viewpoint diversity · metric substitution